1 /*
   2  * Copyright (c) 2012, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "gc/shared/barrierSet.hpp"
  27 #include "opto/arraycopynode.hpp"
  28 #include "oops/objArrayKlass.hpp"
  29 #include "opto/convertnode.hpp"
  30 #include "opto/graphKit.hpp"
  31 #include "opto/macro.hpp"
  32 #include "opto/runtime.hpp"
  33 #include "utilities/align.hpp"
  34 #if INCLUDE_SHENANDOAHGC
  35 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  36 #include "gc/shenandoah/shenandoahRuntime.hpp"
  37 #endif
  38 
  39 
  40 void PhaseMacroExpand::insert_mem_bar(Node** ctrl, Node** mem, int opcode, Node* precedent) {
  41   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
  42   mb->init_req(TypeFunc::Control, *ctrl);
  43   mb->init_req(TypeFunc::Memory, *mem);
  44   transform_later(mb);
  45   *ctrl = new ProjNode(mb,TypeFunc::Control);
  46   transform_later(*ctrl);
  47   Node* mem_proj = new ProjNode(mb,TypeFunc::Memory);
  48   transform_later(mem_proj);
  49   *mem = mem_proj;
  50 }
  51 
  52 Node* PhaseMacroExpand::array_element_address(Node* ary, Node* idx, BasicType elembt) {
  53   uint shift  = exact_log2(type2aelembytes(elembt));
  54   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
  55   Node* base =  basic_plus_adr(ary, header);
  56 #ifdef _LP64
  57   // see comment in GraphKit::array_element_address
  58   int index_max = max_jint - 1;  // array size is max_jint, index is one less
  59   const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
  60   idx = transform_later( new ConvI2LNode(idx, lidxtype) );
  61 #endif
  62   Node* scale = new LShiftXNode(idx, intcon(shift));
  63   transform_later(scale);
  64   return basic_plus_adr(ary, base, scale);
  65 }
  66 
  67 Node* PhaseMacroExpand::ConvI2L(Node* offset) {
  68   return transform_later(new ConvI2LNode(offset));
  69 }
  70 
  71 Node* PhaseMacroExpand::make_leaf_call(Node* ctrl, Node* mem,
  72                                        const TypeFunc* call_type, address call_addr,
  73                                        const char* call_name,
  74                                        const TypePtr* adr_type,
  75                                        Node* parm0, Node* parm1,
  76                                        Node* parm2, Node* parm3,
  77                                        Node* parm4, Node* parm5,
  78                                        Node* parm6, Node* parm7) {
  79   int size = call_type->domain()->cnt();
  80   Node* call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
  81   call->init_req(TypeFunc::Control, ctrl);
  82   call->init_req(TypeFunc::I_O    , top());
  83   call->init_req(TypeFunc::Memory , mem);
  84   call->init_req(TypeFunc::ReturnAdr, top());
  85   call->init_req(TypeFunc::FramePtr, top());
  86 
  87   // Hook each parm in order.  Stop looking at the first NULL.
  88   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
  89   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
  90   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
  91   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
  92   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
  93   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
  94   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
  95   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
  96     /* close each nested if ===> */  } } } } } } } }
  97   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
  98 
  99   return call;
 100 }
 101 
 102 
 103 //------------------------------generate_guard---------------------------
 104 // Helper function for generating guarded fast-slow graph structures.
 105 // The given 'test', if true, guards a slow path.  If the test fails
 106 // then a fast path can be taken.  (We generally hope it fails.)
 107 // In all cases, GraphKit::control() is updated to the fast path.
 108 // The returned value represents the control for the slow path.
 109 // The return value is never 'top'; it is either a valid control
 110 // or NULL if it is obvious that the slow path can never be taken.
 111 // Also, if region and the slow control are not NULL, the slow edge
 112 // is appended to the region.
 113 Node* PhaseMacroExpand::generate_guard(Node** ctrl, Node* test, RegionNode* region, float true_prob) {
 114   if ((*ctrl)->is_top()) {
 115     // Already short circuited.
 116     return NULL;
 117   }
 118   // Build an if node and its projections.
 119   // If test is true we take the slow path, which we assume is uncommon.
 120   if (_igvn.type(test) == TypeInt::ZERO) {
 121     // The slow branch is never taken.  No need to build this guard.
 122     return NULL;
 123   }
 124 
 125   IfNode* iff = new IfNode(*ctrl, test, true_prob, COUNT_UNKNOWN);
 126   transform_later(iff);
 127 
 128   Node* if_slow = new IfTrueNode(iff);
 129   transform_later(if_slow);
 130 
 131   if (region != NULL) {
 132     region->add_req(if_slow);
 133   }
 134 
 135   Node* if_fast = new IfFalseNode(iff);
 136   transform_later(if_fast);
 137 
 138   *ctrl = if_fast;
 139 
 140   return if_slow;
 141 }
 142 
 143 inline Node* PhaseMacroExpand::generate_slow_guard(Node** ctrl, Node* test, RegionNode* region) {
 144   return generate_guard(ctrl, test, region, PROB_UNLIKELY_MAG(3));
 145 }
 146 
 147 void PhaseMacroExpand::generate_negative_guard(Node** ctrl, Node* index, RegionNode* region) {
 148   if ((*ctrl)->is_top())
 149     return;                // already stopped
 150   if (_igvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
 151     return;                // index is already adequately typed
 152   Node* cmp_lt = new CmpINode(index, intcon(0));
 153   transform_later(cmp_lt);
 154   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 155   transform_later(bol_lt);
 156   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 157 }
 158 
 159 void PhaseMacroExpand::generate_limit_guard(Node** ctrl, Node* offset, Node* subseq_length, Node* array_length, RegionNode* region) {
 160   if ((*ctrl)->is_top())
 161     return;                // already stopped
 162   bool zero_offset = _igvn.type(offset) == TypeInt::ZERO;
 163   if (zero_offset && subseq_length->eqv_uncast(array_length))
 164     return;                // common case of whole-array copy
 165   Node* last = subseq_length;
 166   if (!zero_offset) {            // last += offset
 167     last = new AddINode(last, offset);
 168     transform_later(last);
 169   }
 170   Node* cmp_lt = new CmpUNode(array_length, last);
 171   transform_later(cmp_lt);
 172   Node* bol_lt = new BoolNode(cmp_lt, BoolTest::lt);
 173   transform_later(bol_lt);
 174   generate_guard(ctrl, bol_lt, region, PROB_MIN);
 175 }
 176 
 177 Node* PhaseMacroExpand::generate_nonpositive_guard(Node** ctrl, Node* index, bool never_negative) {
 178   if ((*ctrl)->is_top())  return NULL;
 179 
 180   if (_igvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
 181     return NULL;                // index is already adequately typed
 182   Node* cmp_le = new CmpINode(index, intcon(0));
 183   transform_later(cmp_le);
 184   BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
 185   Node* bol_le = new BoolNode(cmp_le, le_or_eq);
 186   transform_later(bol_le);
 187   Node* is_notp = generate_guard(ctrl, bol_le, NULL, PROB_MIN);
 188 
 189   return is_notp;
 190 }
 191 
 192 void PhaseMacroExpand::finish_arraycopy_call(Node* call, Node** ctrl, MergeMemNode** mem, const TypePtr* adr_type) {
 193   transform_later(call);
 194 
 195   *ctrl = new ProjNode(call,TypeFunc::Control);
 196   transform_later(*ctrl);
 197   Node* newmem = new ProjNode(call, TypeFunc::Memory);
 198   transform_later(newmem);
 199 
 200   uint alias_idx = C->get_alias_index(adr_type);
 201   if (alias_idx != Compile::AliasIdxBot) {
 202     *mem = MergeMemNode::make(*mem);
 203     (*mem)->set_memory_at(alias_idx, newmem);
 204   } else {
 205     *mem = MergeMemNode::make(newmem);
 206   }
 207   transform_later(*mem);
 208 }
 209 
 210 address PhaseMacroExpand::basictype2arraycopy(BasicType t,
 211                                               Node* src_offset,
 212                                               Node* dest_offset,
 213                                               bool disjoint_bases,
 214                                               const char* &name,
 215                                               bool dest_uninitialized) {
 216   const TypeInt* src_offset_inttype  = _igvn.find_int_type(src_offset);;
 217   const TypeInt* dest_offset_inttype = _igvn.find_int_type(dest_offset);;
 218 
 219   bool aligned = false;
 220   bool disjoint = disjoint_bases;
 221 
 222   // if the offsets are the same, we can treat the memory regions as
 223   // disjoint, because either the memory regions are in different arrays,
 224   // or they are identical (which we can treat as disjoint.)  We can also
 225   // treat a copy with a destination index  less that the source index
 226   // as disjoint since a low->high copy will work correctly in this case.
 227   if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
 228       dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
 229     // both indices are constants
 230     int s_offs = src_offset_inttype->get_con();
 231     int d_offs = dest_offset_inttype->get_con();
 232     int element_size = type2aelembytes(t);
 233     aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
 234               ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
 235     if (s_offs >= d_offs)  disjoint = true;
 236   } else if (src_offset == dest_offset && src_offset != NULL) {
 237     // This can occur if the offsets are identical non-constants.
 238     disjoint = true;
 239   }
 240 
 241   return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
 242 }
 243 
 244 #define XTOP LP64_ONLY(COMMA top())
 245 
 246 // Generate an optimized call to arraycopy.
 247 // Caller must guard against non-arrays.
 248 // Caller must determine a common array basic-type for both arrays.
 249 // Caller must validate offsets against array bounds.
 250 // The slow_region has already collected guard failure paths
 251 // (such as out of bounds length or non-conformable array types).
 252 // The generated code has this shape, in general:
 253 //
 254 //     if (length == 0)  return   // via zero_path
 255 //     slowval = -1
 256 //     if (types unknown) {
 257 //       slowval = call generic copy loop
 258 //       if (slowval == 0)  return  // via checked_path
 259 //     } else if (indexes in bounds) {
 260 //       if ((is object array) && !(array type check)) {
 261 //         slowval = call checked copy loop
 262 //         if (slowval == 0)  return  // via checked_path
 263 //       } else {
 264 //         call bulk copy loop
 265 //         return  // via fast_path
 266 //       }
 267 //     }
 268 //     // adjust params for remaining work:
 269 //     if (slowval != -1) {
 270 //       n = -1^slowval; src_offset += n; dest_offset += n; length -= n
 271 //     }
 272 //   slow_region:
 273 //     call slow arraycopy(src, src_offset, dest, dest_offset, length)
 274 //     return  // via slow_call_path
 275 //
 276 // This routine is used from several intrinsics:  System.arraycopy,
 277 // Object.clone (the array subcase), and Arrays.copyOf[Range].
 278 //
 279 Node* PhaseMacroExpand::generate_arraycopy(ArrayCopyNode *ac, AllocateArrayNode* alloc,
 280                                            Node** ctrl, MergeMemNode* mem, Node** io,
 281                                            const TypePtr* adr_type,
 282                                            BasicType basic_elem_type,
 283                                            Node* src,  Node* src_offset,
 284                                            Node* dest, Node* dest_offset,
 285                                            Node* copy_length,
 286                                            bool disjoint_bases,
 287                                            bool length_never_negative,
 288                                            RegionNode* slow_region) {
 289   if (slow_region == NULL) {
 290     slow_region = new RegionNode(1);
 291     transform_later(slow_region);
 292   }
 293 
 294   Node* original_dest      = dest;
 295   bool  dest_uninitialized = false;
 296 
 297   // See if this is the initialization of a newly-allocated array.
 298   // If so, we will take responsibility here for initializing it to zero.
 299   // (Note:  Because tightly_coupled_allocation performs checks on the
 300   // out-edges of the dest, we need to avoid making derived pointers
 301   // from it until we have checked its uses.)
 302   if (ReduceBulkZeroing
 303       && !(UseTLAB && ZeroTLAB) // pointless if already zeroed
 304       && basic_elem_type != T_CONFLICT // avoid corner case
 305       && !src->eqv_uncast(dest)
 306       && alloc != NULL
 307       && _igvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
 308       && alloc->maybe_set_complete(&_igvn)) {
 309     // "You break it, you buy it."
 310     InitializeNode* init = alloc->initialization();
 311     assert(init->is_complete(), "we just did this");
 312     init->set_complete_with_arraycopy();
 313     assert(dest->is_CheckCastPP(), "sanity");
 314     assert(dest->in(0)->in(0) == init, "dest pinned");
 315     adr_type = TypeRawPtr::BOTTOM;  // all initializations are into raw memory
 316     // From this point on, every exit path is responsible for
 317     // initializing any non-copied parts of the object to zero.
 318     // Also, if this flag is set we make sure that arraycopy interacts properly
 319     // with G1, eliding pre-barriers. See CR 6627983.
 320     dest_uninitialized = true;
 321   } else {
 322     // No zeroing elimination here.
 323     alloc             = NULL;
 324     //original_dest   = dest;
 325     //dest_uninitialized = false;
 326   }
 327 
 328   uint alias_idx = C->get_alias_index(adr_type);
 329 
 330   // Results are placed here:
 331   enum { fast_path        = 1,  // normal void-returning assembly stub
 332          checked_path     = 2,  // special assembly stub with cleanup
 333          slow_call_path   = 3,  // something went wrong; call the VM
 334          zero_path        = 4,  // bypass when length of copy is zero
 335          bcopy_path       = 5,  // copy primitive array by 64-bit blocks
 336          PATH_LIMIT       = 6
 337   };
 338   RegionNode* result_region = new RegionNode(PATH_LIMIT);
 339   PhiNode*    result_i_o    = new PhiNode(result_region, Type::ABIO);
 340   PhiNode*    result_memory = new PhiNode(result_region, Type::MEMORY, adr_type);
 341   assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
 342   transform_later(result_region);
 343   transform_later(result_i_o);
 344   transform_later(result_memory);
 345 
 346   // The slow_control path:
 347   Node* slow_control;
 348   Node* slow_i_o = *io;
 349   Node* slow_mem = mem->memory_at(alias_idx);
 350   DEBUG_ONLY(slow_control = (Node*) badAddress);
 351 
 352   // Checked control path:
 353   Node* checked_control = top();
 354   Node* checked_mem     = NULL;
 355   Node* checked_i_o     = NULL;
 356   Node* checked_value   = NULL;
 357 
 358   if (basic_elem_type == T_CONFLICT) {
 359     assert(!dest_uninitialized, "");
 360     Node* cv = generate_generic_arraycopy(ctrl, &mem,
 361                                           adr_type,
 362                                           src, src_offset, dest, dest_offset,
 363                                           copy_length, dest_uninitialized);
 364     if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
 365     checked_control = *ctrl;
 366     checked_i_o     = *io;
 367     checked_mem     = mem->memory_at(alias_idx);
 368     checked_value   = cv;
 369     *ctrl = top();
 370   }
 371 
 372   Node* not_pos = generate_nonpositive_guard(ctrl, copy_length, length_never_negative);
 373   if (not_pos != NULL) {
 374     Node* local_ctrl = not_pos, *local_io = *io;
 375     MergeMemNode* local_mem = MergeMemNode::make(mem);
 376     transform_later(local_mem);
 377 
 378     // (6) length must not be negative.
 379     if (!length_never_negative) {
 380       generate_negative_guard(&local_ctrl, copy_length, slow_region);
 381     }
 382 
 383     // copy_length is 0.
 384     if (dest_uninitialized) {
 385       assert(!local_ctrl->is_top(), "no ctrl?");
 386       Node* dest_length = alloc->in(AllocateNode::ALength);
 387       if (copy_length->eqv_uncast(dest_length)
 388           || _igvn.find_int_con(dest_length, 1) <= 0) {
 389         // There is no zeroing to do. No need for a secondary raw memory barrier.
 390       } else {
 391         // Clear the whole thing since there are no source elements to copy.
 392         generate_clear_array(local_ctrl, local_mem,
 393                              adr_type, dest, basic_elem_type,
 394                              intcon(0), NULL,
 395                              alloc->in(AllocateNode::AllocSize));
 396         // Use a secondary InitializeNode as raw memory barrier.
 397         // Currently it is needed only on this path since other
 398         // paths have stub or runtime calls as raw memory barriers.
 399         MemBarNode* mb = MemBarNode::make(C, Op_Initialize,
 400                                           Compile::AliasIdxRaw,
 401                                           top());
 402         transform_later(mb);
 403         mb->set_req(TypeFunc::Control,local_ctrl);
 404         mb->set_req(TypeFunc::Memory, local_mem->memory_at(Compile::AliasIdxRaw));
 405         local_ctrl = transform_later(new ProjNode(mb, TypeFunc::Control));
 406         local_mem->set_memory_at(Compile::AliasIdxRaw, transform_later(new ProjNode(mb, TypeFunc::Memory)));
 407 
 408         InitializeNode* init = mb->as_Initialize();
 409         init->set_complete(&_igvn);  // (there is no corresponding AllocateNode)
 410       }
 411     }
 412 
 413     // Present the results of the fast call.
 414     result_region->init_req(zero_path, local_ctrl);
 415     result_i_o   ->init_req(zero_path, local_io);
 416     result_memory->init_req(zero_path, local_mem->memory_at(alias_idx));
 417   }
 418 
 419   if (!(*ctrl)->is_top() && dest_uninitialized) {
 420     // We have to initialize the *uncopied* part of the array to zero.
 421     // The copy destination is the slice dest[off..off+len].  The other slices
 422     // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
 423     Node* dest_size   = alloc->in(AllocateNode::AllocSize);
 424     Node* dest_length = alloc->in(AllocateNode::ALength);
 425     Node* dest_tail   = transform_later( new AddINode(dest_offset, copy_length));
 426 
 427     // If there is a head section that needs zeroing, do it now.
 428     if (_igvn.find_int_con(dest_offset, -1) != 0) {
 429       generate_clear_array(*ctrl, mem,
 430                            adr_type, dest, basic_elem_type,
 431                            intcon(0), dest_offset,
 432                            NULL);
 433     }
 434 
 435     // Next, perform a dynamic check on the tail length.
 436     // It is often zero, and we can win big if we prove this.
 437     // There are two wins:  Avoid generating the ClearArray
 438     // with its attendant messy index arithmetic, and upgrade
 439     // the copy to a more hardware-friendly word size of 64 bits.
 440     Node* tail_ctl = NULL;
 441     if (!(*ctrl)->is_top() && !dest_tail->eqv_uncast(dest_length)) {
 442       Node* cmp_lt   = transform_later( new CmpINode(dest_tail, dest_length) );
 443       Node* bol_lt   = transform_later( new BoolNode(cmp_lt, BoolTest::lt) );
 444       tail_ctl = generate_slow_guard(ctrl, bol_lt, NULL);
 445       assert(tail_ctl != NULL || !(*ctrl)->is_top(), "must be an outcome");
 446     }
 447 
 448     // At this point, let's assume there is no tail.
 449     if (!(*ctrl)->is_top() && alloc != NULL && basic_elem_type != T_OBJECT) {
 450       // There is no tail.  Try an upgrade to a 64-bit copy.
 451       bool didit = false;
 452       {
 453         Node* local_ctrl = *ctrl, *local_io = *io;
 454         MergeMemNode* local_mem = MergeMemNode::make(mem);
 455         transform_later(local_mem);
 456 
 457         didit = generate_block_arraycopy(&local_ctrl, &local_mem, local_io,
 458                                          adr_type, basic_elem_type, alloc,
 459                                          src, src_offset, dest, dest_offset,
 460                                          dest_size, dest_uninitialized);
 461         if (didit) {
 462           // Present the results of the block-copying fast call.
 463           result_region->init_req(bcopy_path, local_ctrl);
 464           result_i_o   ->init_req(bcopy_path, local_io);
 465           result_memory->init_req(bcopy_path, local_mem->memory_at(alias_idx));
 466         }
 467       }
 468       if (didit) {
 469         *ctrl = top();     // no regular fast path
 470       }
 471     }
 472 
 473     // Clear the tail, if any.
 474     if (tail_ctl != NULL) {
 475       Node* notail_ctl = (*ctrl)->is_top() ? NULL : *ctrl;
 476       *ctrl = tail_ctl;
 477       if (notail_ctl == NULL) {
 478         generate_clear_array(*ctrl, mem,
 479                              adr_type, dest, basic_elem_type,
 480                              dest_tail, NULL,
 481                              dest_size);
 482       } else {
 483         // Make a local merge.
 484         Node* done_ctl = transform_later(new RegionNode(3));
 485         Node* done_mem = transform_later(new PhiNode(done_ctl, Type::MEMORY, adr_type));
 486         done_ctl->init_req(1, notail_ctl);
 487         done_mem->init_req(1, mem->memory_at(alias_idx));
 488         generate_clear_array(*ctrl, mem,
 489                              adr_type, dest, basic_elem_type,
 490                              dest_tail, NULL,
 491                              dest_size);
 492         done_ctl->init_req(2, *ctrl);
 493         done_mem->init_req(2, mem->memory_at(alias_idx));
 494         *ctrl = done_ctl;
 495         mem->set_memory_at(alias_idx, done_mem);
 496       }
 497     }
 498   }
 499 
 500   BasicType copy_type = basic_elem_type;
 501   assert(basic_elem_type != T_ARRAY, "caller must fix this");
 502   if (!(*ctrl)->is_top() && copy_type == T_OBJECT) {
 503     // If src and dest have compatible element types, we can copy bits.
 504     // Types S[] and D[] are compatible if D is a supertype of S.
 505     //
 506     // If they are not, we will use checked_oop_disjoint_arraycopy,
 507     // which performs a fast optimistic per-oop check, and backs off
 508     // further to JVM_ArrayCopy on the first per-oop check that fails.
 509     // (Actually, we don't move raw bits only; the GC requires card marks.)
 510 
 511     // We don't need a subtype check for validated copies and Object[].clone()
 512     bool skip_subtype_check = ac->is_arraycopy_validated() || ac->is_copyof_validated() ||
 513                               ac->is_copyofrange_validated() || ac->is_cloneoop();
 514     if (!skip_subtype_check) {
 515       // Get the klass* for both src and dest
 516       Node* src_klass  = ac->in(ArrayCopyNode::SrcKlass);
 517       Node* dest_klass = ac->in(ArrayCopyNode::DestKlass);
 518 
 519       assert(src_klass != NULL && dest_klass != NULL, "should have klasses");
 520 
 521       // Generate the subtype check.
 522       // This might fold up statically, or then again it might not.
 523       //
 524       // Non-static example:  Copying List<String>.elements to a new String[].
 525       // The backing store for a List<String> is always an Object[],
 526       // but its elements are always type String, if the generic types
 527       // are correct at the source level.
 528       //
 529       // Test S[] against D[], not S against D, because (probably)
 530       // the secondary supertype cache is less busy for S[] than S.
 531       // This usually only matters when D is an interface.
 532       Node* not_subtype_ctrl = Phase::gen_subtype_check(src_klass, dest_klass, ctrl, mem, &_igvn);
 533       // Plug failing path into checked_oop_disjoint_arraycopy
 534       if (not_subtype_ctrl != top()) {
 535         Node* local_ctrl = not_subtype_ctrl;
 536         MergeMemNode* local_mem = MergeMemNode::make(mem);
 537         transform_later(local_mem);
 538 
 539         // (At this point we can assume disjoint_bases, since types differ.)
 540         int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
 541         Node* p1 = basic_plus_adr(dest_klass, ek_offset);
 542         Node* n1 = LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), p1, TypeRawPtr::BOTTOM);
 543         Node* dest_elem_klass = transform_later(n1);
 544         Node* cv = generate_checkcast_arraycopy(&local_ctrl, &local_mem,
 545                                                 adr_type,
 546                                                 dest_elem_klass,
 547                                                 src, src_offset, dest, dest_offset,
 548                                                 ConvI2X(copy_length), dest_uninitialized);
 549         if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
 550         checked_control = local_ctrl;
 551         checked_i_o     = *io;
 552         checked_mem     = local_mem->memory_at(alias_idx);
 553         checked_value   = cv;
 554       }
 555     }
 556     // At this point we know we do not need type checks on oop stores.
 557 
 558     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 559     if (alloc != NULL && !bs->array_copy_requires_gc_barriers(copy_type) && !UseShenandoahGC) {
 560       // If we do not need gc barriers, copy using the jint or jlong stub.
 561       copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
 562       assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
 563              "sizes agree");
 564     }
 565   }
 566 
 567   if (!(*ctrl)->is_top()) {
 568     // Generate the fast path, if possible.
 569     Node* local_ctrl = *ctrl;
 570     MergeMemNode* local_mem = MergeMemNode::make(mem);
 571     transform_later(local_mem);
 572 
 573     generate_unchecked_arraycopy(&local_ctrl, &local_mem,
 574                                  adr_type, copy_type, disjoint_bases,
 575                                  src, src_offset, dest, dest_offset,
 576                                  ConvI2X(copy_length), dest_uninitialized);
 577 
 578     // Present the results of the fast call.
 579     result_region->init_req(fast_path, local_ctrl);
 580     result_i_o   ->init_req(fast_path, *io);
 581     result_memory->init_req(fast_path, local_mem->memory_at(alias_idx));
 582   }
 583 
 584   // Here are all the slow paths up to this point, in one bundle:
 585   assert(slow_region != NULL, "allocated on entry");
 586   slow_control = slow_region;
 587   DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
 588 
 589   *ctrl = checked_control;
 590   if (!(*ctrl)->is_top()) {
 591     // Clean up after the checked call.
 592     // The returned value is either 0 or -1^K,
 593     // where K = number of partially transferred array elements.
 594     Node* cmp = new CmpINode(checked_value, intcon(0));
 595     transform_later(cmp);
 596     Node* bol = new BoolNode(cmp, BoolTest::eq);
 597     transform_later(bol);
 598     IfNode* iff = new IfNode(*ctrl, bol, PROB_MAX, COUNT_UNKNOWN);
 599     transform_later(iff);
 600 
 601     // If it is 0, we are done, so transfer to the end.
 602     Node* checks_done = new IfTrueNode(iff);
 603     transform_later(checks_done);
 604     result_region->init_req(checked_path, checks_done);
 605     result_i_o   ->init_req(checked_path, checked_i_o);
 606     result_memory->init_req(checked_path, checked_mem);
 607 
 608     // If it is not zero, merge into the slow call.
 609     *ctrl = new IfFalseNode(iff);
 610     transform_later(*ctrl);
 611     RegionNode* slow_reg2 = new RegionNode(3);
 612     PhiNode*    slow_i_o2 = new PhiNode(slow_reg2, Type::ABIO);
 613     PhiNode*    slow_mem2 = new PhiNode(slow_reg2, Type::MEMORY, adr_type);
 614     transform_later(slow_reg2);
 615     transform_later(slow_i_o2);
 616     transform_later(slow_mem2);
 617     slow_reg2  ->init_req(1, slow_control);
 618     slow_i_o2  ->init_req(1, slow_i_o);
 619     slow_mem2  ->init_req(1, slow_mem);
 620     slow_reg2  ->init_req(2, *ctrl);
 621     slow_i_o2  ->init_req(2, checked_i_o);
 622     slow_mem2  ->init_req(2, checked_mem);
 623 
 624     slow_control = slow_reg2;
 625     slow_i_o     = slow_i_o2;
 626     slow_mem     = slow_mem2;
 627 
 628     if (alloc != NULL) {
 629       // We'll restart from the very beginning, after zeroing the whole thing.
 630       // This can cause double writes, but that's OK since dest is brand new.
 631       // So we ignore the low 31 bits of the value returned from the stub.
 632     } else {
 633       // We must continue the copy exactly where it failed, or else
 634       // another thread might see the wrong number of writes to dest.
 635       Node* checked_offset = new XorINode(checked_value, intcon(-1));
 636       Node* slow_offset    = new PhiNode(slow_reg2, TypeInt::INT);
 637       transform_later(checked_offset);
 638       transform_later(slow_offset);
 639       slow_offset->init_req(1, intcon(0));
 640       slow_offset->init_req(2, checked_offset);
 641 
 642       // Adjust the arguments by the conditionally incoming offset.
 643       Node* src_off_plus  = new AddINode(src_offset,  slow_offset);
 644       transform_later(src_off_plus);
 645       Node* dest_off_plus = new AddINode(dest_offset, slow_offset);
 646       transform_later(dest_off_plus);
 647       Node* length_minus  = new SubINode(copy_length, slow_offset);
 648       transform_later(length_minus);
 649 
 650       // Tweak the node variables to adjust the code produced below:
 651       src_offset  = src_off_plus;
 652       dest_offset = dest_off_plus;
 653       copy_length = length_minus;
 654     }
 655   }
 656   *ctrl = slow_control;
 657   if (!(*ctrl)->is_top()) {
 658     Node* local_ctrl = *ctrl, *local_io = slow_i_o;
 659     MergeMemNode* local_mem = MergeMemNode::make(mem);
 660     transform_later(local_mem);
 661 
 662     // Generate the slow path, if needed.
 663     local_mem->set_memory_at(alias_idx, slow_mem);
 664 
 665     if (dest_uninitialized) {
 666       generate_clear_array(local_ctrl, local_mem,
 667                            adr_type, dest, basic_elem_type,
 668                            intcon(0), NULL,
 669                            alloc->in(AllocateNode::AllocSize));
 670     }
 671 
 672     local_mem = generate_slow_arraycopy(ac,
 673                                         &local_ctrl, local_mem, &local_io,
 674                                         adr_type,
 675                                         src, src_offset, dest, dest_offset,
 676                                         copy_length, /*dest_uninitialized*/false);
 677 
 678     result_region->init_req(slow_call_path, local_ctrl);
 679     result_i_o   ->init_req(slow_call_path, local_io);
 680     result_memory->init_req(slow_call_path, local_mem->memory_at(alias_idx));
 681   } else {
 682     ShouldNotReachHere(); // no call to generate_slow_arraycopy:
 683                           // projections were not extracted
 684   }
 685 
 686   // Remove unused edges.
 687   for (uint i = 1; i < result_region->req(); i++) {
 688     if (result_region->in(i) == NULL) {
 689       result_region->init_req(i, top());
 690     }
 691   }
 692 
 693   // Finished; return the combined state.
 694   *ctrl = result_region;
 695   *io = result_i_o;
 696   mem->set_memory_at(alias_idx, result_memory);
 697 
 698   // mem no longer guaranteed to stay a MergeMemNode
 699   Node* out_mem = mem;
 700   DEBUG_ONLY(mem = NULL);
 701 
 702   // The memory edges above are precise in order to model effects around
 703   // array copies accurately to allow value numbering of field loads around
 704   // arraycopy.  Such field loads, both before and after, are common in Java
 705   // collections and similar classes involving header/array data structures.
 706   //
 707   // But with low number of register or when some registers are used or killed
 708   // by arraycopy calls it causes registers spilling on stack. See 6544710.
 709   // The next memory barrier is added to avoid it. If the arraycopy can be
 710   // optimized away (which it can, sometimes) then we can manually remove
 711   // the membar also.
 712   //
 713   // Do not let reads from the cloned object float above the arraycopy.
 714   if (alloc != NULL && !alloc->initialization()->does_not_escape()) {
 715     // Do not let stores that initialize this object be reordered with
 716     // a subsequent store that would make this object accessible by
 717     // other threads.
 718     insert_mem_bar(ctrl, &out_mem, Op_MemBarStoreStore);
 719   } else if (InsertMemBarAfterArraycopy) {
 720     insert_mem_bar(ctrl, &out_mem, Op_MemBarCPUOrder);
 721   }
 722 
 723   _igvn.replace_node(_memproj_fallthrough, out_mem);
 724   _igvn.replace_node(_ioproj_fallthrough, *io);
 725   _igvn.replace_node(_fallthroughcatchproj, *ctrl);
 726 
 727 #ifdef ASSERT
 728   const TypeOopPtr* dest_t = _igvn.type(dest)->is_oopptr();
 729   if (dest_t->is_known_instance()) {
 730     ArrayCopyNode* ac = NULL;
 731     assert(ArrayCopyNode::may_modify(dest_t, (*ctrl)->in(0)->as_MemBar(), &_igvn, ac), "dependency on arraycopy lost");
 732     assert(ac == NULL, "no arraycopy anymore");
 733   }
 734 #endif
 735 
 736   return out_mem;
 737 }
 738 
 739 // Helper for initialization of arrays, creating a ClearArray.
 740 // It writes zero bits in [start..end), within the body of an array object.
 741 // The memory effects are all chained onto the 'adr_type' alias category.
 742 //
 743 // Since the object is otherwise uninitialized, we are free
 744 // to put a little "slop" around the edges of the cleared area,
 745 // as long as it does not go back into the array's header,
 746 // or beyond the array end within the heap.
 747 //
 748 // The lower edge can be rounded down to the nearest jint and the
 749 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
 750 //
 751 // Arguments:
 752 //   adr_type           memory slice where writes are generated
 753 //   dest               oop of the destination array
 754 //   basic_elem_type    element type of the destination
 755 //   slice_idx          array index of first element to store
 756 //   slice_len          number of elements to store (or NULL)
 757 //   dest_size          total size in bytes of the array object
 758 //
 759 // Exactly one of slice_len or dest_size must be non-NULL.
 760 // If dest_size is non-NULL, zeroing extends to the end of the object.
 761 // If slice_len is non-NULL, the slice_idx value must be a constant.
 762 void PhaseMacroExpand::generate_clear_array(Node* ctrl, MergeMemNode* merge_mem,
 763                                             const TypePtr* adr_type,
 764                                             Node* dest,
 765                                             BasicType basic_elem_type,
 766                                             Node* slice_idx,
 767                                             Node* slice_len,
 768                                             Node* dest_size) {
 769   // one or the other but not both of slice_len and dest_size:
 770   assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
 771   if (slice_len == NULL)  slice_len = top();
 772   if (dest_size == NULL)  dest_size = top();
 773 
 774   uint alias_idx = C->get_alias_index(adr_type);
 775 
 776   // operate on this memory slice:
 777   Node* mem = merge_mem->memory_at(alias_idx); // memory slice to operate on
 778 
 779   // scaling and rounding of indexes:
 780   int scale = exact_log2(type2aelembytes(basic_elem_type));
 781   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 782   int clear_low = (-1 << scale) & (BytesPerInt  - 1);
 783   int bump_bit  = (-1 << scale) & BytesPerInt;
 784 
 785   // determine constant starts and ends
 786   const intptr_t BIG_NEG = -128;
 787   assert(BIG_NEG + 2*abase < 0, "neg enough");
 788   intptr_t slice_idx_con = (intptr_t) _igvn.find_int_con(slice_idx, BIG_NEG);
 789   intptr_t slice_len_con = (intptr_t) _igvn.find_int_con(slice_len, BIG_NEG);
 790   if (slice_len_con == 0) {
 791     return;                     // nothing to do here
 792   }
 793   intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
 794   intptr_t end_con   = _igvn.find_intptr_t_con(dest_size, -1);
 795   if (slice_idx_con >= 0 && slice_len_con >= 0) {
 796     assert(end_con < 0, "not two cons");
 797     end_con = align_up(abase + ((slice_idx_con + slice_len_con) << scale),
 798                        BytesPerLong);
 799   }
 800 
 801   if (start_con >= 0 && end_con >= 0) {
 802     // Constant start and end.  Simple.
 803     mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
 804                                        start_con, end_con, &_igvn);
 805   } else if (start_con >= 0 && dest_size != top()) {
 806     // Constant start, pre-rounded end after the tail of the array.
 807     Node* end = dest_size;
 808     mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
 809                                        start_con, end, &_igvn);
 810   } else if (start_con >= 0 && slice_len != top()) {
 811     // Constant start, non-constant end.  End needs rounding up.
 812     // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
 813     intptr_t end_base  = abase + (slice_idx_con << scale);
 814     int      end_round = (-1 << scale) & (BytesPerLong  - 1);
 815     Node*    end       = ConvI2X(slice_len);
 816     if (scale != 0)
 817       end = transform_later(new LShiftXNode(end, intcon(scale) ));
 818     end_base += end_round;
 819     end = transform_later(new AddXNode(end, MakeConX(end_base)) );
 820     end = transform_later(new AndXNode(end, MakeConX(~end_round)) );
 821     mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
 822                                        start_con, end, &_igvn);
 823   } else if (start_con < 0 && dest_size != top()) {
 824     // Non-constant start, pre-rounded end after the tail of the array.
 825     // This is almost certainly a "round-to-end" operation.
 826     Node* start = slice_idx;
 827     start = ConvI2X(start);
 828     if (scale != 0)
 829       start = transform_later(new LShiftXNode( start, intcon(scale) ));
 830     start = transform_later(new AddXNode(start, MakeConX(abase)) );
 831     if ((bump_bit | clear_low) != 0) {
 832       int to_clear = (bump_bit | clear_low);
 833       // Align up mod 8, then store a jint zero unconditionally
 834       // just before the mod-8 boundary.
 835       if (((abase + bump_bit) & ~to_clear) - bump_bit
 836           < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
 837         bump_bit = 0;
 838         assert((abase & to_clear) == 0, "array base must be long-aligned");
 839       } else {
 840         // Bump 'start' up to (or past) the next jint boundary:
 841         start = transform_later( new AddXNode(start, MakeConX(bump_bit)) );
 842         assert((abase & clear_low) == 0, "array base must be int-aligned");
 843       }
 844       // Round bumped 'start' down to jlong boundary in body of array.
 845       start = transform_later(new AndXNode(start, MakeConX(~to_clear)) );
 846       if (bump_bit != 0) {
 847         // Store a zero to the immediately preceding jint:
 848         Node* x1 = transform_later(new AddXNode(start, MakeConX(-bump_bit)) );
 849         Node* p1 = basic_plus_adr(dest, x1);
 850         mem = StoreNode::make(_igvn, ctrl, mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
 851         mem = transform_later(mem);
 852       }
 853     }
 854     Node* end = dest_size; // pre-rounded
 855     mem = ClearArrayNode::clear_memory(ctrl, mem, dest,
 856                                        start, end, &_igvn);
 857   } else {
 858     // Non-constant start, unrounded non-constant end.
 859     // (Nobody zeroes a random midsection of an array using this routine.)
 860     ShouldNotReachHere();       // fix caller
 861   }
 862 
 863   // Done.
 864   merge_mem->set_memory_at(alias_idx, mem);
 865 }
 866 
 867 bool PhaseMacroExpand::generate_block_arraycopy(Node** ctrl, MergeMemNode** mem, Node* io,
 868                                                 const TypePtr* adr_type,
 869                                                 BasicType basic_elem_type,
 870                                                 AllocateNode* alloc,
 871                                                 Node* src,  Node* src_offset,
 872                                                 Node* dest, Node* dest_offset,
 873                                                 Node* dest_size, bool dest_uninitialized) {
 874   // See if there is an advantage from block transfer.
 875   int scale = exact_log2(type2aelembytes(basic_elem_type));
 876   if (scale >= LogBytesPerLong)
 877     return false;               // it is already a block transfer
 878 
 879   // Look at the alignment of the starting offsets.
 880   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
 881 
 882   intptr_t src_off_con  = (intptr_t) _igvn.find_int_con(src_offset, -1);
 883   intptr_t dest_off_con = (intptr_t) _igvn.find_int_con(dest_offset, -1);
 884   if (src_off_con < 0 || dest_off_con < 0) {
 885     // At present, we can only understand constants.
 886     return false;
 887   }
 888 
 889   intptr_t src_off  = abase + (src_off_con  << scale);
 890   intptr_t dest_off = abase + (dest_off_con << scale);
 891 
 892   if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
 893     // Non-aligned; too bad.
 894     // One more chance:  Pick off an initial 32-bit word.
 895     // This is a common case, since abase can be odd mod 8.
 896     if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
 897         ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
 898       Node* sptr = basic_plus_adr(src,  src_off);
 899       Node* dptr = basic_plus_adr(dest, dest_off);
 900       const TypePtr* s_adr_type = _igvn.type(sptr)->is_ptr();
 901       assert(s_adr_type->isa_aryptr(), "impossible slice");
 902       uint s_alias_idx = C->get_alias_index(s_adr_type);
 903       uint d_alias_idx = C->get_alias_index(adr_type);
 904       bool is_mismatched = (basic_elem_type != T_INT);
 905       Node* sval = transform_later(
 906           LoadNode::make(_igvn, *ctrl, (*mem)->memory_at(s_alias_idx), sptr, s_adr_type,
 907                          TypeInt::INT, T_INT, MemNode::unordered, LoadNode::DependsOnlyOnTest,
 908                          false /*unaligned*/, is_mismatched));
 909       Node* st = transform_later(
 910           StoreNode::make(_igvn, *ctrl, (*mem)->memory_at(d_alias_idx), dptr, adr_type,
 911                           sval, T_INT, MemNode::unordered));
 912       if (is_mismatched) {
 913         st->as_Store()->set_mismatched_access();
 914       }
 915       (*mem)->set_memory_at(d_alias_idx, st);
 916       src_off += BytesPerInt;
 917       dest_off += BytesPerInt;
 918     } else {
 919       return false;
 920     }
 921   }
 922   assert(src_off % BytesPerLong == 0, "");
 923   assert(dest_off % BytesPerLong == 0, "");
 924 
 925   // Do this copy by giant steps.
 926   Node* sptr  = basic_plus_adr(src,  src_off);
 927   Node* dptr  = basic_plus_adr(dest, dest_off);
 928   Node* countx = dest_size;
 929   countx = transform_later(new SubXNode(countx, MakeConX(dest_off)));
 930   countx = transform_later(new URShiftXNode(countx, intcon(LogBytesPerLong)));
 931 
 932   bool disjoint_bases = true;   // since alloc != NULL
 933   generate_unchecked_arraycopy(ctrl, mem,
 934                                adr_type, T_LONG, disjoint_bases,
 935                                sptr, NULL, dptr, NULL, countx, dest_uninitialized);
 936 
 937   return true;
 938 }
 939 
 940 // Helper function; generates code for the slow case.
 941 // We make a call to a runtime method which emulates the native method,
 942 // but without the native wrapper overhead.
 943 MergeMemNode* PhaseMacroExpand::generate_slow_arraycopy(ArrayCopyNode *ac,
 944                                                         Node** ctrl, Node* mem, Node** io,
 945                                                         const TypePtr* adr_type,
 946                                                         Node* src,  Node* src_offset,
 947                                                         Node* dest, Node* dest_offset,
 948                                                         Node* copy_length, bool dest_uninitialized) {
 949   assert(!dest_uninitialized, "Invariant");
 950 
 951   const TypeFunc* call_type = OptoRuntime::slow_arraycopy_Type();
 952   CallNode* call = new CallStaticJavaNode(call_type, OptoRuntime::slow_arraycopy_Java(),
 953                                           "slow_arraycopy",
 954                                           ac->jvms()->bci(), TypePtr::BOTTOM);
 955 
 956   call->init_req(TypeFunc::Control, *ctrl);
 957   call->init_req(TypeFunc::I_O    , *io);
 958   call->init_req(TypeFunc::Memory , mem);
 959   call->init_req(TypeFunc::ReturnAdr, top());
 960   call->init_req(TypeFunc::FramePtr, top());
 961   call->init_req(TypeFunc::Parms+0, src);
 962   call->init_req(TypeFunc::Parms+1, src_offset);
 963   call->init_req(TypeFunc::Parms+2, dest);
 964   call->init_req(TypeFunc::Parms+3, dest_offset);
 965   call->init_req(TypeFunc::Parms+4, copy_length);
 966   copy_call_debug_info(ac, call);
 967 
 968   call->set_cnt(PROB_UNLIKELY_MAG(4));  // Same effect as RC_UNCOMMON.
 969   _igvn.replace_node(ac, call);
 970   transform_later(call);
 971 
 972   extract_call_projections(call);
 973   *ctrl = _fallthroughcatchproj->clone();
 974   transform_later(*ctrl);
 975 
 976   Node* m = _memproj_fallthrough->clone();
 977   transform_later(m);
 978 
 979   uint alias_idx = C->get_alias_index(adr_type);
 980   MergeMemNode* out_mem;
 981   if (alias_idx != Compile::AliasIdxBot) {
 982     out_mem = MergeMemNode::make(mem);
 983     out_mem->set_memory_at(alias_idx, m);
 984   } else {
 985     out_mem = MergeMemNode::make(m);
 986   }
 987   transform_later(out_mem);
 988 
 989   *io = _ioproj_fallthrough->clone();
 990   transform_later(*io);
 991 
 992   return out_mem;
 993 }
 994 
 995 // Helper function; generates code for cases requiring runtime checks.
 996 Node* PhaseMacroExpand::generate_checkcast_arraycopy(Node** ctrl, MergeMemNode** mem,
 997                                                      const TypePtr* adr_type,
 998                                                      Node* dest_elem_klass,
 999                                                      Node* src,  Node* src_offset,
1000                                                      Node* dest, Node* dest_offset,
1001                                                      Node* copy_length, bool dest_uninitialized) {
1002   if ((*ctrl)->is_top())  return NULL;
1003 
1004   address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
1005   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1006     return NULL;
1007   }
1008 
1009   // Pick out the parameters required to perform a store-check
1010   // for the target array.  This is an optimistic check.  It will
1011   // look in each non-null element's class, at the desired klass's
1012   // super_check_offset, for the desired klass.
1013   int sco_offset = in_bytes(Klass::super_check_offset_offset());
1014   Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
1015   Node* n3 = new LoadINode(NULL, *mem /*memory(p3)*/, p3, _igvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
1016   Node* check_offset = ConvI2X(transform_later(n3));
1017   Node* check_value  = dest_elem_klass;
1018 
1019   Node* src_start  = array_element_address(src,  src_offset,  T_OBJECT);
1020   Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
1021 
1022   const TypeFunc* call_type = OptoRuntime::checkcast_arraycopy_Type();
1023   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "checkcast_arraycopy", adr_type,
1024                               src_start, dest_start, copy_length XTOP, check_offset XTOP, check_value);
1025 
1026   finish_arraycopy_call(call, ctrl, mem, adr_type);
1027 
1028   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1029   transform_later(proj);
1030 
1031   return proj;
1032 }
1033 
1034 // Helper function; generates code for cases requiring runtime checks.
1035 Node* PhaseMacroExpand::generate_generic_arraycopy(Node** ctrl, MergeMemNode** mem,
1036                                                    const TypePtr* adr_type,
1037                                                    Node* src,  Node* src_offset,
1038                                                    Node* dest, Node* dest_offset,
1039                                                    Node* copy_length, bool dest_uninitialized) {
1040   if ((*ctrl)->is_top()) return NULL;
1041   assert(!dest_uninitialized, "Invariant");
1042 
1043   address copyfunc_addr = StubRoutines::generic_arraycopy();
1044   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
1045     return NULL;
1046   }
1047 
1048   const TypeFunc* call_type = OptoRuntime::generic_arraycopy_Type();
1049   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, "generic_arraycopy", adr_type,
1050                               src, src_offset, dest, dest_offset, copy_length);
1051 
1052   finish_arraycopy_call(call, ctrl, mem, adr_type);
1053 
1054   Node* proj =  new ProjNode(call, TypeFunc::Parms);
1055   transform_later(proj);
1056 
1057   return proj;
1058 }
1059 
1060 // Helper function; generates the fast out-of-line call to an arraycopy stub.
1061 void PhaseMacroExpand::generate_unchecked_arraycopy(Node** ctrl, MergeMemNode** mem,
1062                                                     const TypePtr* adr_type,
1063                                                     BasicType basic_elem_type,
1064                                                     bool disjoint_bases,
1065                                                     Node* src,  Node* src_offset,
1066                                                     Node* dest, Node* dest_offset,
1067                                                     Node* copy_length, bool dest_uninitialized) {
1068   if ((*ctrl)->is_top()) return;
1069 
1070   Node* src_start  = src;
1071   Node* dest_start = dest;
1072   if (src_offset != NULL || dest_offset != NULL) {
1073     src_start =  array_element_address(src, src_offset, basic_elem_type);
1074     dest_start = array_element_address(dest, dest_offset, basic_elem_type);
1075   }
1076 
1077   // Figure out which arraycopy runtime method to call.
1078   const char* copyfunc_name = "arraycopy";
1079   address     copyfunc_addr =
1080       basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
1081                           disjoint_bases, copyfunc_name, dest_uninitialized);
1082 
1083   const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1084   Node* call = make_leaf_call(*ctrl, *mem, call_type, copyfunc_addr, copyfunc_name, adr_type,
1085                               src_start, dest_start, copy_length XTOP);
1086 
1087   finish_arraycopy_call(call, ctrl, mem, adr_type);
1088 }
1089 
1090 #if INCLUDE_SHENANDOAHGC
1091 Node* PhaseMacroExpand::shenandoah_call_clone_barrier(Node* call, Node* dest) {
1092   assert (UseShenandoahGC && ShenandoahCloneBarrier, "Should be enabled");
1093   const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
1094   Node* c = new ProjNode(call,TypeFunc::Control);
1095   transform_later(c);
1096   Node* m = new ProjNode(call, TypeFunc::Memory);
1097   transform_later(m);
1098   assert(dest->is_AddP(), "bad input");
1099   call = make_leaf_call(c, m, ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
1100                         CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
1101                         "shenandoah_clone_barrier", raw_adr_type, dest->in(AddPNode::Base));
1102   transform_later(call);
1103   return call;
1104 }
1105 #endif
1106 
1107 void PhaseMacroExpand::expand_arraycopy_node(ArrayCopyNode *ac) {
1108   Node* ctrl = ac->in(TypeFunc::Control);
1109   Node* io = ac->in(TypeFunc::I_O);
1110   Node* src = ac->in(ArrayCopyNode::Src);
1111   Node* src_offset = ac->in(ArrayCopyNode::SrcPos);
1112   Node* dest = ac->in(ArrayCopyNode::Dest);
1113   Node* dest_offset = ac->in(ArrayCopyNode::DestPos);
1114   Node* length = ac->in(ArrayCopyNode::Length);
1115   MergeMemNode* merge_mem = NULL;
1116 
1117   if (ac->is_clonebasic()) {
1118     assert (src_offset == NULL && dest_offset == NULL, "for clone offsets should be null");
1119     Node* mem = ac->in(TypeFunc::Memory);
1120     const char* copyfunc_name = "arraycopy";
1121     address     copyfunc_addr =
1122       basictype2arraycopy(T_LONG, NULL, NULL,
1123                           true, copyfunc_name, true);
1124 
1125     const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
1126     const TypeFunc* call_type = OptoRuntime::fast_arraycopy_Type();
1127 
1128     Node* call = make_leaf_call(ctrl, mem, call_type, copyfunc_addr, copyfunc_name, raw_adr_type, src, dest, length XTOP);
1129     transform_later(call);
1130 
1131 #if INCLUDE_SHENANDOAHGC
1132     if (UseShenandoahGC && ShenandoahCloneBarrier) {
1133       const TypeOopPtr* src_type = _igvn.type(src)->is_oopptr();
1134       if (src_type->isa_instptr() != NULL) {
1135         ciInstanceKlass* ik = src_type->klass()->as_instance_klass();
1136         if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) {
1137           if (ik->has_object_fields()) {
1138             call = shenandoah_call_clone_barrier(call, dest);
1139           } else {
1140             if (!src_type->klass_is_exact()) {
1141               C->dependencies()->assert_leaf_type(ik);
1142             }
1143           }
1144         } else {
1145           call = shenandoah_call_clone_barrier(call, dest);
1146         }
1147       } else if (src_type->isa_aryptr()) {
1148         BasicType src_elem  = src_type->klass()->as_array_klass()->element_type()->basic_type();
1149         if (src_elem == T_OBJECT || src_elem == T_ARRAY) {
1150           call = shenandoah_call_clone_barrier(call, dest);
1151         }
1152       } else {
1153         call = shenandoah_call_clone_barrier(call, dest);
1154       }
1155     }
1156 #endif
1157 
1158     _igvn.replace_node(ac, call);
1159     return;
1160   } else if (ac->is_copyof() || ac->is_copyofrange() || ac->is_cloneoop()) {
1161     Node* mem = ac->in(TypeFunc::Memory);
1162     merge_mem = MergeMemNode::make(mem);
1163     transform_later(merge_mem);
1164 
1165     RegionNode* slow_region = new RegionNode(1);
1166     transform_later(slow_region);
1167 
1168     AllocateArrayNode* alloc = NULL;
1169     if (ac->is_alloc_tightly_coupled()) {
1170       alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1171       assert(alloc != NULL, "expect alloc");
1172     }
1173 
1174     const TypePtr* adr_type = _igvn.type(dest)->is_oopptr()->add_offset(Type::OffsetBot);
1175     if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1176       adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1177     }
1178     generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1179                        adr_type, T_OBJECT,
1180                        src, src_offset, dest, dest_offset, length,
1181                        true, !ac->is_copyofrange());
1182 
1183     return;
1184   }
1185 
1186   AllocateArrayNode* alloc = NULL;
1187   if (ac->is_alloc_tightly_coupled()) {
1188     alloc = AllocateArrayNode::Ideal_array_allocation(dest, &_igvn);
1189     assert(alloc != NULL, "expect alloc");
1190   }
1191 
1192   assert(ac->is_arraycopy() || ac->is_arraycopy_validated(), "should be an arraycopy");
1193 
1194   // Compile time checks.  If any of these checks cannot be verified at compile time,
1195   // we do not make a fast path for this call.  Instead, we let the call remain as it
1196   // is.  The checks we choose to mandate at compile time are:
1197   //
1198   // (1) src and dest are arrays.
1199   const Type* src_type = src->Value(&_igvn);
1200   const Type* dest_type = dest->Value(&_igvn);
1201   const TypeAryPtr* top_src = src_type->isa_aryptr();
1202   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
1203 
1204   BasicType src_elem = T_CONFLICT;
1205   BasicType dest_elem = T_CONFLICT;
1206 
1207   if (top_dest != NULL && top_dest->klass() != NULL) {
1208     dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
1209   }
1210   if (top_src != NULL && top_src->klass() != NULL) {
1211     src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
1212   }
1213   if (src_elem  == T_ARRAY)  src_elem  = T_OBJECT;
1214   if (dest_elem == T_ARRAY)  dest_elem = T_OBJECT;
1215 
1216   if (ac->is_arraycopy_validated() &&
1217       dest_elem != T_CONFLICT &&
1218       src_elem == T_CONFLICT) {
1219     src_elem = dest_elem;
1220   }
1221 
1222   if (src_elem == T_CONFLICT || dest_elem == T_CONFLICT) {
1223     // Conservatively insert a memory barrier on all memory slices.
1224     // Do not let writes into the source float below the arraycopy.
1225     {
1226       Node* mem = ac->in(TypeFunc::Memory);
1227       insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1228 
1229       merge_mem = MergeMemNode::make(mem);
1230       transform_later(merge_mem);
1231     }
1232 
1233     // Call StubRoutines::generic_arraycopy stub.
1234     Node* mem = generate_arraycopy(ac, NULL, &ctrl, merge_mem, &io,
1235                                    TypeRawPtr::BOTTOM, T_CONFLICT,
1236                                    src, src_offset, dest, dest_offset, length,
1237                                    // If a  negative length guard was generated for the ArrayCopyNode,
1238                                    // the length of the array can never be negative.
1239                                    false, ac->has_negative_length_guard());
1240 
1241     // Do not let reads from the destination float above the arraycopy.
1242     // Since we cannot type the arrays, we don't know which slices
1243     // might be affected.  We could restrict this barrier only to those
1244     // memory slices which pertain to array elements--but don't bother.
1245     if (!InsertMemBarAfterArraycopy) {
1246       // (If InsertMemBarAfterArraycopy, there is already one in place.)
1247       insert_mem_bar(&ctrl, &mem, Op_MemBarCPUOrder);
1248     }
1249     return;
1250   }
1251 
1252   assert(!ac->is_arraycopy_validated() || (src_elem == dest_elem && dest_elem != T_VOID), "validated but different basic types");
1253 
1254   // (2) src and dest arrays must have elements of the same BasicType
1255   // Figure out the size and type of the elements we will be copying.
1256   if (src_elem != dest_elem || dest_elem == T_VOID) {
1257     // The component types are not the same or are not recognized.  Punt.
1258     // (But, avoid the native method wrapper to JVM_ArrayCopy.)
1259     {
1260       Node* mem = ac->in(TypeFunc::Memory);
1261       merge_mem = generate_slow_arraycopy(ac, &ctrl, mem, &io, TypePtr::BOTTOM, src, src_offset, dest, dest_offset, length, false);
1262     }
1263 
1264     _igvn.replace_node(_memproj_fallthrough, merge_mem);
1265     _igvn.replace_node(_ioproj_fallthrough, io);
1266     _igvn.replace_node(_fallthroughcatchproj, ctrl);
1267     return;
1268   }
1269 
1270   //---------------------------------------------------------------------------
1271   // We will make a fast path for this call to arraycopy.
1272 
1273   // We have the following tests left to perform:
1274   //
1275   // (3) src and dest must not be null.
1276   // (4) src_offset must not be negative.
1277   // (5) dest_offset must not be negative.
1278   // (6) length must not be negative.
1279   // (7) src_offset + length must not exceed length of src.
1280   // (8) dest_offset + length must not exceed length of dest.
1281   // (9) each element of an oop array must be assignable
1282 
1283   {
1284     Node* mem = ac->in(TypeFunc::Memory);
1285     merge_mem = MergeMemNode::make(mem);
1286     transform_later(merge_mem);
1287   }
1288 
1289   RegionNode* slow_region = new RegionNode(1);
1290   transform_later(slow_region);
1291 
1292   if (!ac->is_arraycopy_validated()) {
1293     // (3) operands must not be null
1294     // We currently perform our null checks with the null_check routine.
1295     // This means that the null exceptions will be reported in the caller
1296     // rather than (correctly) reported inside of the native arraycopy call.
1297     // This should be corrected, given time.  We do our null check with the
1298     // stack pointer restored.
1299     // null checks done library_call.cpp
1300 
1301     // (4) src_offset must not be negative.
1302     generate_negative_guard(&ctrl, src_offset, slow_region);
1303 
1304     // (5) dest_offset must not be negative.
1305     generate_negative_guard(&ctrl, dest_offset, slow_region);
1306 
1307     // (6) length must not be negative (moved to generate_arraycopy()).
1308     // generate_negative_guard(length, slow_region);
1309 
1310     // (7) src_offset + length must not exceed length of src.
1311     Node* alen = ac->in(ArrayCopyNode::SrcLen);
1312     assert(alen != NULL, "need src len");
1313     generate_limit_guard(&ctrl,
1314                          src_offset, length,
1315                          alen,
1316                          slow_region);
1317 
1318     // (8) dest_offset + length must not exceed length of dest.
1319     alen = ac->in(ArrayCopyNode::DestLen);
1320     assert(alen != NULL, "need dest len");
1321     generate_limit_guard(&ctrl,
1322                          dest_offset, length,
1323                          alen,
1324                          slow_region);
1325 
1326     // (9) each element of an oop array must be assignable
1327     // The generate_arraycopy subroutine checks this.
1328   }
1329   // This is where the memory effects are placed:
1330   const TypePtr* adr_type = NULL;
1331   if (ac->_dest_type != TypeOopPtr::BOTTOM) {
1332     adr_type = ac->_dest_type->add_offset(Type::OffsetBot)->is_ptr();
1333   } else {
1334     adr_type = TypeAryPtr::get_array_body_type(dest_elem);
1335   }
1336 
1337   generate_arraycopy(ac, alloc, &ctrl, merge_mem, &io,
1338                      adr_type, dest_elem,
1339                      src, src_offset, dest, dest_offset, length,
1340                      // If a  negative length guard was generated for the ArrayCopyNode,
1341                      // the length of the array can never be negative.
1342                      false, ac->has_negative_length_guard(), slow_region);
1343 }